On the Feasibility of Fidelity$^-$ for Graph Pruning

Fidelity metrics may not reliably guide graph pruning: Despite fidelity being widely used to evaluate GNN explanations based on the intuitive principle that removing "unimportant" features should minimally affect predictions, the paper questions whether this metric actually works well for the practical task of pruning graphs to improve model efficiency.

General explainability methods outperform GNN-specific approaches: Across testing 7 different edge attribution methods, the researchers found that general explainable AI techniques performed better than methods specifically designed for graph neural networks when used for graph pruning—suggesting that specialized GNN explanation methods may not translate to practical benefits.

Local explanations don't automatically create effective global pruning strategies: The paper reveals a gap between how well local explanations work (explaining individual predictions) and whether they can be aggregated into global masks that effectively prune entire graphs while maintaining model performance—an important distinction for practitioners relying on explanation methods.

Current explanation quality metrics may not align with practical goals: The findings suggest that metrics commonly used to validate explanation quality (like fidelity) may not correlate with real-world objectives like model efficiency, highlighting a potential misalignment between how we evaluate AI explanations and how we should use them in practice.